AI news, November 21, 2026: Riemann claim with Claude, Cursor and Grok Build, Hermes Enterprise, Perplexity Computer
Short answer: The main stories of November 20 to 21: an unconfirmed claim of a proof of the Riemann hypothesis with Claude, an unconfirmed leak about renaming Cursor to Grok Build, Microsoft building on Hermes and OpenClaw, Perplexity Computer as a browser extension and OpenAI research against distillation. Leaked code, paste-images and earnings claims around Grok Bot are unconfirmed.

Illustration by Claude (AI), drawn with code: a wave curve with a magnifying glass (Riemann claim, unconfirmed), two editor windows with an arrow (leak about Cursor and Grok Build) and a browser with an agent cursor (Perplexity Computer).
Riemann hypothesis with Claude: a claim, not yet confirmedConfidence: low for the new claim: X posts, verification under way; high for the earlier Anthropic result: Anthropic's own page.
@kimmonismus reports that Anthropic mathematician Levent Alpöge writes on his website that he has found a proof of the Riemann hypothesis together with Claude, with the proof online as Lean code (607 thousand views, more than nineteen times the account's median). @DeItaone followed with 184 thousand views. Several accounts report that independent verification is still under way. We did not open Alpöge's website or the Lean code ourselves in this round.
Earlier this summer Anthropic itself published a result on the Riemann zeta function: a higher lower bound for the share of zeros on the critical line, explicitly not a proof of the hypothesis. The present claim goes further and is known only through X posts. A video of the Anthropic founders also appeared (295 thousand views), the day after the listing; background on the figures is in our article on Anthropic's S-1/A.
What this means: call this a claim until independent mathematicians have confirmed the proof. A Lean proof that compiles is strong evidence, but whether the checked statement is the Riemann hypothesis is part of that review. Do not adopt it as fact.
Sources: @kimmonismus over de Riemann-claim op X · @DeItaone over de Riemann-claim op X · @Relox_AI over de lopende verificatie op X · @AnthropicAI: video van de oprichters op X · Anthropic: eerder resultaat over de Riemann-zeta-functie
Leak: Cursor may become Grok Build (unconfirmed)Confidence: low: a leak from one account with little verification; no official confirmation.
@GrokInsider writes that Cursor is now called Grok Build, with a new name and a faster app (150 thousand views). @blankspeaker says internal nightly macOS builds have already been renamed and that the Cursor CLI 2026.11.19 contains code that downloads Grok Build, under the name Grok Build 2.0 (Unified). An analysis account itself calls the renaming unconfirmed. We have not seen an official announcement from Cursor or xAI; @cursor_ai posted nothing original in the past 48 hours.
What this means: anyone with scripts that call the cursor command can list those calls now and change them only after an official statement. About the existing Grok Build apps we wrote in our article on Grok Build, and about an earlier security incident at Cursor in our article on the Fireworks key.
Sources: @GrokInsider over Cursor en Grok Build op X · @blankspeaker over hernoemde nightly-builds op X · @blankspeaker over Grok Build 2.0 (Unified) op X · @AGTPinsights: nog onbevestigd, op X
Hermes Enterprise and OpenClaw: Microsoft builds alongConfidence: medium to high: official accounts; the contributor count comes from a single post.
Nous Research's announcement of Hermes Enterprise in Microsoft Teams, the Microsoft Store and Azure reached 682 thousand views, nearly nine times the account's median. The winners of the Plugin Hackathon (almost a thousand entries) are known: agents scored the entries and five community juries chose the top 10. @openclaw reports that Microsoft is also building on OpenClaw and supplying maintainers (15 thousand views); @Teknium reports more than 5,000 contributors for Hermes.
What this means: backing from a large company is a signal of project health. Anyone who wants to use community plugins from the hackathon reads the code first. We covered the announcement itself in our article on Hermes Enterprise in Microsoft Teams, and the approval step in OpenClaw 2026.11.4.
Sources: @NousResearch over Hermes Enterprise op X · @NousResearch: winnaars Plugin Hackathon op X · @openclaw over Microsoft en OpenClaw op X · @Teknium over het aantal contributors op X
Perplexity Computer works as an extension in Chromium browsersConfidence: medium to high for the extension: official account; low for Gemini 4 Neon as default: a consideration, not a decision.
Perplexity reports that Computer now runs as an extension in Chromium browsers: you hand over a task from the side panel and Computer works across your tabs and tools (31 thousand views). @AravSrinivas is considering Gemini 4 Neon as the default model in Computer (17.7 thousand views) and mentions a Mac app with both frontier and local models (131 thousand views).
What this means: a browser agent with access to logged-in tabs calls for a separate browser profile without your sensitive sessions. We have not tested the Mac app with local models. About Perplexity's earlier work we wrote in our article on Perplexity Brain and skills, and on Perplexity Teammates.
Sources: @perplexity_ai over Computer als extensie op X · @AravSrinivas over Gemini 4 Neon in Computer op X · @AravSrinivas over de Mac-app met lokale modellen op X
OpenAI publishes research against distillation and introduces MedChemBenchConfidence: high: official accounts; the consequences for users are our interpretation.
@OpenAI publishes research on how capabilities transfer between models and how the company detects coordinated distillation attempts (761 thousand views), and introduces MedChemBench, a benchmark for reasoning in medicinal chemistry (304 thousand views). @OpenAIDevs describes Ultrafast with GPT-6.2 Sol on ChatGPT Pro 200 (50 thousand views). According to OpenAI, companies outside the United States can now claim their business and publish agents in ChatGPT.
What this means: anyone who stores model output at scale can expect stricter detection in the API; the post names no concrete rules. About voice in the API we wrote in our article on GPT-Live-1.1, and about misuse detection at OpenAI and Anthropic in our article on misuse detection.
Sources: @OpenAI over onderzoek tegen distillatie op X · @OpenAI over MedChemBench op X · @OpenAIDevs over Ultrafast met GPT-6.2 Sol op X · @OpenAI over bedrijven buiten de VS op X
Grok Bot with QuickBooks, Cloudflare Browser Run and PlannotatorConfidence: medium to high: official accounts; not tested by us.
Grok Bot connects to Intuit QuickBooks, in the United States only: ask for cash flow and have invoices, quotes and reminders drafted for you to approve (171 thousand views). @CloudflareDev reports that Browser Run /snapshot returns PDF, links, scrape and JSON from a single page load (17 thousand views). Plannotator is an open-source message board for agents on Cloudflare (18 thousand views), and Cloudflare announces new WAF rules against phpAfterShell and ruby2shell, also for the Free plan.
What this means: with Grok Bot the approval stays with the user. Anyone considering Browser Run or Plannotator reads the documentation and the code first; we did not test them. The WAF rules do not replace patching PHP or Ruby software.
Sources: @bot over de QuickBooks-koppeling op X · @CloudflareDev over Browser Run /snapshot op X · @plannotator over het berichtenbord voor agents op X · @Cloudflare over de WAF-regels op X
Unconfirmed and warnings: leaked code, paste-images, 532 plugins and earnings claimsConfidence: low: single posts without a primary source; included as a warning.
One account claims that Grok Bot's source code has leaked (not verified). Two accounts urge people to paste images containing instructions into Grok Bot for 24/7 trading bots (37 and 24 thousand views); images with instructions are a known route for prompt injection. Another spreads an install guide for an uncensored agent on your own devices (32 thousand views). The story about 532 free Claude plugins from Anthropic is circulating again (216 thousand views) and is still not confirmed. Earnings claims with dollar amounts come from accounts showing the same pattern.
What this means: do not paste prompts or images of unknown origin into an agent, do not open leaked code and do not install a plugin without reading its code. About agents with an approval step we wrote in our article on the approval step in OpenClaw 2026.11.4.
Sources: @choopyplug1 over gelekte Grok Bot-code op X · @RohOnChain over plak-afbeeldingen op X · @0xCarnagee over een ongecensureerde agent op X · @PaulYoungX over 532 plugins op X
Primary sources: Anthropic: eerder resultaat over de Riemann-zeta-functie · Perplexity: Computer als extensie (X) · OpenAI: onderzoek tegen distillatie (X)


